cxas-sim-eval
Pass
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests external data from evaluation JSON files and conversation traces, which are interpolated into prompts for the Gemini model during the conversion and diagnostic stages.
- Ingestion points: Evaluation data is read from JSON files in
scripts/convert_eval.pyand conversation logs are parsed inscripts/run_evals.py. - Boundary markers: The prompt templates for both conversion and analysis do not utilize explicit delimiters or instructions to prevent the model from following commands embedded within the input data.
- Capability inventory: The skill scripts have permissions to perform file system operations, call Vertex AI Gemini APIs, and upload content to Google Cloud Storage.
- Sanitization: No evidence of sanitization, validation, or escaping of the ingested data was found before it is passed to the language model.
Audit Metadata